A ComfyUI SDXL workflow for photorealistic images should make the lighting, subject and camera perspective agree. Adding “ultra realistic” repeatedly cannot compensate for a hand that bends incorrectly or a face lit from two conflicting directions.
Establish a model-compatible baseline
The project’s official SDXL examples describe using the base checkpoint and working around 1024×1024 pixels or comparable pixel counts at other aspect ratios. They also provide base-plus-refiner examples. Start with the simplest applicable graph, then evaluate whether an extra stage improves your deliverable.
If you use an SDXL-derived checkpoint such as a particular Juggernaut release, read that release’s model card. A setting copied from a different checkpoint or version is a starting hypothesis, not a universal preset. Record the exact model filename with your graph.
Write a photographic brief
Use a compact description with five parts: subject, location, framing, light and action. For an original adult creator, a starting brief could be: “Chest-up editorial portrait of an adult travel host in a navy jacket, standing beside a station window, soft daylight from camera left, relaxed expression, simple background.”
This is an original example to adapt, not a tested preset. It avoids asking the model to solve a crowded scene, complex hand gesture and extreme camera angle simultaneously. Once the baseline passes, add one of those challenges deliberately.
Evaluate the image in layers
| Layer | What to check |
|---|---|
| Composition | The crop supports the intended post and leaves room for text |
| Light | Highlights and shadows agree across the face, clothing and room |
| Anatomy | Hands, teeth, ears and joints remain plausible |
| Material | Fabric, skin, glass and metal do not share the same artificial texture |
| Identity | The creator matches the approved reference, if this is a recurring persona |
A small prompt experiment
Keep one prompt and one model fixed. First compare a simple front-facing portrait with a three-quarter pose. Then compare the same pose under soft window light and shaded outdoor light. Label each result and reject any image that introduces new identity or anatomy problems.
This sequence gives you a reason for each revision. If changing the lighting improves texture but changes the apparent identity, note the conflict. Do not replace the reference image with whichever generation happens to look most attractive.
Use refiners and repairs only when they solve a named problem
A refiner or second pass adds another opportunity to change the image. Compare before and after crops at the same size. Keep the additional stage only if it improves the accepted asset without introducing a new failure.
For a small defect, use a bounded inpainting repair. For a larger final file, use a separate upscale test. A clear sequence—generate, review, repair, review, enlarge—makes it easier to identify where an unwanted change entered the image.
Prepare an image for later animation
Choose a pose that can support the intended motion. A hand already hidden behind a complex object is a poor starting point for a large wave. Keep the face readable, inspect the edges of the body and leave enough space for the action you plan to request.
Save the exact approved still beside the video brief. The Wan image-to-video prompt guide shows how to turn that still into a simple motion instruction.
Choose the production environment that fits your work
Use ComfyUI when inspecting and maintaining the graph is part of your process. If the aim is recurring creator content in a hosted studio, build an original persona in Clout. In either environment, judge the final asset against the same written brief and identity reference.
